Graph Learning · LLM × Graph · Multi-Agent · Science
Showing 3 papers for 2026-10-10
Nothing cleared the bar today. Papers read in full today: 2.
AgentGarten provides a framework that couples simulators and game engines with a shared neural renderer to create real-time interactive environments with persistent world state. The system addresses the bottleneck of achieving both fidelity and realism across diverse worlds by integrating simulation backends with neural rendering. Its simulation backends maintain persistent world state and execute program-defined dynamics.
Learn2Play Bench introduces a benchmark of newly designed text-based games to evaluate how well LLM agents learn from experience in unfamiliar environments, separating learning from interaction from reasoning with existing knowledge. Existing benchmarks often test tasks whose rules are provided or already familiar, making it hard to distinguish learning from interactions from reasoning. To address this, Learn2Play Bench presents novel text-based challenges that probe learning in novel rules and dynamics.
TokenRouter proposes an efficient serving system for token-level routing of LLMs to improve efficiency and quality. While coarse-grained routing at the session or query level is widely adopted, fine-grained token-level routing can yield substantial gains. However, efficiently serving token-level routed inference poses significant challenges, and current systems built on single-LLM assumptions suffer from step desynchronization and frequent batch admission issues.
Trace2Env explores agentic language world modeling by using a world-modeling agent as the environment for a task agent, enabling faithful, stateful simulations from historical traces when the original environment is unavailable. It proposes a learning-free framework that reconstructs traces into reusable simulations, preserving behavioral fidelity without re-creating the executable system.
SuperNav introduces an agentic navigation system for any task in any scene, aiming for general-purpose navigation for service robots. Some existing methods fine-tune multimodal LLMs to predict navigation actions, making behavior dependent on training data coverage and limiting generalization. The key idea is to let the MLLM interpret requests, understand scenes, and make decisions while delegating low-level control to specialized modules, preserving generality.
Finsler Flow Matching introduces a dynamics-aware geodesic interpolation for inferring cellular trajectories from single-snapshot data. It goes beyond transport inferred from population marginals by encoding directional information with a cell-cell Markov transition kernel and a Finsler-geometric formulation to interpolate along flow-aware geodesics. The approach aims to produce more accurate trajectory inferences by matching dynamic flows rather than relying solely on symmetric transport geometry.
Looking Inside LLMs investigates whether internal network organization constrains or enables reasoning by measuring small-world connectivity in graphs built from attention-head activation similarities. By constructing functional graphs from attention-head activations, the authors show that a higher small-world index correlates with better reasoning performance, suggesting SWI captures efficient local clustering and short global paths in reasoning networks. These results imply that internal architectural organization of LLMs can be a signature of reasoning capacity beyond standard behavioral metrics.
This paper presents DAG-EDA, an interactive system that lets analysts and an LLM co-navigate the space of possible analyses for exploratory data analysis. Analysts often work from high-level domain questions toward concrete analyses, while LLM responses can be unstructured, obscuring what has been explored and why certain directions were chosen. DAG-EDA uses an Intent Graph to explicitly trace the analytic reasoning path and provide a structured, auditable view of exploration that supports human–machine collaboration.